Patent Analytics at Dtangle

This post gives an overview of my work at Dtangle, a company specialised in patent analytics. Client engagements are confidential, so it describes methods only: no client, invention or result is disclosed.

The problem

Patent decisions (whether to file, what to claim, how much room a portfolio leaves) are usually made on a handful of documents read by hand. At Dtangle, the goal is to replace that with structured evidence: every concept, every prior-art reference and every conclusion is recorded, graded and traceable back to the passages that support it.

What I work on

• Pre-filing analysis. Starting from unpublished ideas, I break them down into individual concepts, write a one-sentence novelty spine for each before any search, then search the patent and scientific literature on generalised vocabulary (never on the confidential wording). The filing structure (how many applications, in which order, what to leave unfiled) is drawn from the concepts that survive.
• Cross-patent dependency analysis. Mapping which concepts of a patent family are carried by earlier or later families, and where a portfolio exposes itself to its own publications as prior art.
• Concept-cluster landscapes. Extracting concepts from patents and publications, clustering them and locating the gaps where nothing has been claimed yet.
• Concept knowledge graphs for technology monitoring. Designing a repository that unifies a curated technology taxonomy, publications (OpenAlex), patents (Lens) and company data around the same concepts, in continuity with my GNN work on technology monitoring.
• Data quality. Assignee disambiguation, auditing open versus proprietary patent data sources, and validation methods to measure how far each source can be trusted.
• Research-funding mapping. Linking public research projects to policy goals (e.g. EU-funded projects to the UN Sustainable Development Goals) through explicit, reviewable mapping rules.

How the work is built

• Analyses are gated: a step only runs once the patent status and the inputs it depends on are established.
• Deliverables (workbooks, reports, executive notes) are generated from the recorded evidence, never edited by hand, so each figure can be audited.
• Confidential invention material never leaves the analysis environment in identifiable form.

Skills involved

Python, data pipelines and CSV/Excel data models, patent and bibliographic databases (Lens, OpenAlex, CORDIS), LLM-assisted concept extraction, and writing decision briefs for non-technical stakeholders.